Globally Consistent Non-rigid Map Fusion for MASt3R-SLAM

Learning based monocular SLAM systems such as MASt3R-SLAM achieve impressive dense reconstructions by leveraging learned priors, but suffer from non-physical deformations such as ghosting and structural misalignment due to inconsistent multi-frame fusion. In this work, we propose a globally consistent non-rigid map fusion framework that addresses these artifacts by introducing edge-guided deformation graphs into the SLAM pipeline. First, we restore metric scale by fusing predicted and measured depth maps via weighted least squares. Then, we extract geometrically meaningful control points using image-edge cues and match them across frames using MASt3R correspondences. These control points guide a deformation graph, where local non-rigid transformations are optimized via a combination of data fidelity, smoothness, and regularization losses. Finally, the deformation field is propagated to non-control points using Gaussian interpolation. Experiments on the Habitat simulator demonstrate that our method significantly reduces ghosting and improves reconstruction accuracy, outperforming existing methods both quantitatively and qualitatively.